Skip to main content

Setting Up Quality Analysis Intelligence (QAI): Best Practices

Best practices for QAI after Granular Control—Control profiles, analysis setup, questions, disposition priority, and testing.

Written by Moja Bot

Quality Analysis Intelligence (QAI) uses AI to automatically analyze your call recordings, extract key information, and classify calls based on custom criteria you define. This guide helps you set up QAI effectively to get accurate, actionable insights.

Before you start

Recommended setup order:

  1. Confirm Moja enabled QAI on your account and you are signed in as an Admin on a full Moja account (publisher-only views don’t include QAI—move to your own full account to unlock it)

  2. Confirm recordings are on for the traffic you’ll analyze

  3. Create/enable a Control profile and attach entities; set sample %

  4. Design dispositions and questions for that profile

  5. Assign disposition priorities (winner among matching labels)

  6. Test on new completed calls; refine

💡 Tip: Start simple with 3–4 core dispositions on one profile at 100% sample while validating. You can always add more later.

See also: QAI Prerequisites: Account, Control, and Recordings and Granular Control release notes.

Step 1: Plan your dispositions

Dispositions are the final classification of a call—the most important outcome that happened.

Common dispositions for most businesses:

  • Sale – Customer made a purchase

  • Callback – Customer requested follow-up

  • Not Interested – Customer declined offer

  • Spam/Scam – Invalid or fraudulent call

To add a disposition:

  1. Go to QAI Analysis → Control

  2. Open the profile that owns this traffic

  3. Open Analysis setup

  4. Click add disposition

  5. Enter name, priority (higher = more important when multiple match), questions, and match rules

  6. Save on the profile

Attaching traffic (which calls get transcribed) is configured on the profile’s Control settings—not inside the disposition form. Campaign-only disposition setup without an enabled Control profile is incomplete after Granular Control.

Step 2: Create effective questions

✅ Do:

  • Ask one thing per question

  • Use language that matches how agents and callers actually speak

  • Prefer concrete outcomes (“Did the customer agree to purchase?”)

❌ Don’t:

  • Stack multiple conditions into one question

  • Use internal jargon callers never say

  • Assume the transcript contains data that was never discussed

❌ Bad (compound question): “Did the customer buy and provide a card and confirm shipping?”

✅ Good (separate questions): “Did the customer agree to buy?” then “Was payment information collected?” then “Were fulfillment details confirmed?”

Step 3: Build questions for key dispositions

Sale / Conversion — Priority 100 (highest)

  • Did the customer agree to purchase or enroll?

  • Was payment or enrollment information collected?

  • Did the agent confirm the order or policy details?

Callback / Follow-up — Priority ~60

  • Did the customer ask to be called back?

  • Was a follow-up time discussed?

Not Interested — Priority ~50

  • Did the customer clearly decline the offer?

  • Did the customer say they are not interested or already covered?

Spam / Invalid — Priority ~10 (lowest among labels you still want captured)

  • Is this a wrong number or non-customer?

  • Does the call appear to be spam, silence, or fraud?

Step 4: Understand disposition priority

When multiple dispositions match a call, Moja selects the one with the highest disposition priority number inside the winning Control profile.

Example: a call matches Sale (100) and Customer Service (60) → final label is Sale.

Priority guidelines:

  • 90–100: revenue outcomes (Sale, Booked, Qualified transfer you monetize)

  • 50–80: meaningful non-sale outcomes (Callback, Not Interested, Existing customer)

  • 1–20: invalid/spam/low-value catch-alls

Leave gaps (100, 80, 60…) so you can insert new labels later without renumbering everything.

Control profile priority vs disposition priority

  • Control profile priority — which profile’s analysis and sample rate apply when multiple enabled profiles match the same call (higher wins). Only one profile wins; QAI does not fan out one call across multiple profiles.

  • Disposition priority — which label wins when multiple dispositions match inside that winning profile.

These are different systems. Do not use one number for both mental models.

Step 5: Choose answer types

  • Yes/No — binary decisions (Did X happen?)

  • Multiple choice — categorizing responses

  • Number — numeric extraction (quote amount, age)

  • Text — short open extraction (order id, reason)

Step 6: Test your setup

Check:

  • Control profile enabled + correct entities attached

  • Sample rate understood (100% while proving the path)

  • Transcript exists on new eligible completes

  • Questions answer from real speech

  • Highest-value disposition wins when several match

Remember: sample rate below 100% intentionally skips many calls. QAI usage follows eligible transcribed/analyzed calls (scope + sample %), not every dial.

Step 7: Refine based on results

  • False positives: tighten wording and match rules

  • Not triggering: confirm transcript and profile analysis setup first; then loosen match criteria / rephrase

  • Wrong winner: adjust disposition priorities

Common mistakes to avoid

  • Skipping Control and only editing campaign dispositions

  • Testing only on old calls from before the profile existed

  • Compound questions

  • Identical priorities on competing high-value labels

  • Disabling recordings on the path under test

Where to review results

  • Call details — transcript + per-question answers + selected disposition

  • QAI reports — trends, filters, exports

Related articles

Did this answer your question?